An Expert System for Automatic Classification of Sound Signals

2020 International Conference on Inventive Computation Technologies (ICICT), 2020

Bird Classifier is a system which will use machine learning approach and allows users to record bird sounds and identify them. The System will help bird species to be identified by taking input only the sound of that bird.The framework also offers tools for translating such annotations into datasets that can be used to train a computer to identify a species ' presence or absence. Dataset recorded are preprocessed to remove unwanted noise and divide useful sounds in frames so that it can act as an input to classifier. The system was tested through different algorithms and the algorithm that gave best results was chosen for implementation.System uses audio features like MFCC,Mel-Spectra etc. This system will use different algorithms such as KNN, Random Forest, Multi layer perceptron, Bayes in classification of birds species.

IEEE Access

Music has been an integral part of the history of humankind with theories suggesting it is more antediluvian than speech itself. Music is an ordered succession of tones and harmonies that produce sounds characterized by melody and rhythm. Our paper proposes an ensemble deep learning musical instrument classification (MIC) framework, named as MIC_FuzzyNET model which aims to classify the dominant instruments present in musical clips. Firstly, the musical data is converted to three different spectrograms: Constant Q-Transform, Semitone Spectrogram, and Mel Spectrogram, which are then stacked to form 3 channel 2D data. This stacked spectrogram is fed to transfer learning models namely, EfficientNetV2 and ResNet18 which output the preliminary classification scores. A fuzzy rank ensemble model is finally employed that assigns the classifier ranks, on the testing data to achieve final enhanced classification scores which reduces error and biases for the constituent CNN architectures. Our proposed framework has been evaluated on the Persian Classical Music Instrument Recognition (PCMIR) dataset and Instrument Recognition in Musical Audio Signals (IRMAS) dataset. It has achieved considerably high accuracy, making our proposed framework a robust MIC model. INDEX TERMS Musical instrument classification, MIC_FuzzyNET, fuzzy integral, spectrogram, transfer learning, PCMIR dataset, IRMAS dataset. I. INTRODUCTION 16 Sound plays an integral role in how living beings perceive 17 the world around them and communicate with each other. 18 Although conscious communication is multi-sensory and 19 involves tactile as well as visual cues in addition to audio 20 The associate editor coordinating the review of this manuscript and approving it for publication was Zahid Akhtar. 25 Audio comes in many forms including random noises, 26 verbal speech, wildlife and environmental sounds, and music, 27 which this paper deals with. Our research proposes an intel-28 VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 197 taken from the dataset to show that our proposed model can 198 perform well even with limited datasets. Furthermore, the 199 5 instruments belong to a different classes of musical instru-200 ments, adding to the robustness of our proposed framework. 201 We have evenly and randomly split the samples in a 8:1:1 202 ratio as shown in Table 2 into training-testing-validation sub-203 sets. This particular ratio was decided to enable the model to 204 train on enough data samples since our proposed framework 205 deals with small datasets while optimizing the number of 206 unseen testing samples on which our framework infers the 207 performance of the ensemble model. The data samples are 208 not collected in a studio environment but collected across 209 different genres, artists, and decades which is why there is 210 great variety in the quality of data points. 211 V. METHODOLOGY 212 The proposed framework has been divided into the follow-213 ing subsections, namely feature extraction, creating stacked 214 spectrogram, CNNs, model training with transfer learning 215 and finally assigning the fuzzy ranks to the CNN models for 216 ensemble learning. 217 A. FEATURE EXTRACTION 218 An eclectic choice of features from the musical samples we 219 have is of prime necessity if we want to extract more informa-220 tion for instrument classification. There is the availability of 221 various temporal and spectral features corresponding to audio 222 data which include Mel spectrograms, LPCC, and MFCC to 223 name a few. However, there are a few spectrograms that are 224 easy to visualize and can be used for our MIC problem. For 225 our proposed model, we have chosen three features which 226 are the CQT spectrogram, Semitone Spectrogram, and Mel 227 Spectrogram. 228 from Ud instrument class of the PCMIR dataset along with 270 its Mel spectrogram which has been extracted using Librosa library in Python. 272 3) SEMITONE SPECTROGRAM The smallest musical interval employed in Western music is 274 the semitone, commonly referred to as a half step or half-275 tone. When performed harmonically, a semitone is thought 276 to be the most discordant [38]. A semitone is the distance 277 in pitch between two notes that are close to one another on 278 a 12-tone scale. When a test signal is run through a signal 279 processor, such as a filter, the results are typically analysed 280 using spectrograms to show the performance. For particular 281 spine notes, the semitones filter determines melodic semitone 282 intervals. The filter can highlight repetitions, steps, leaps, and 283 the direction of intervals in the rendered notation. 284 4) SIGNIFICANCE OF CHOSEN SPECTROGRAMS 285 The hertz values are remapped to the Mel scale in the Mel 286 spectrogram. So, Mel spectrograms are better suited for appli-287 cations that need to replicate human hearing perception, such 288 as music. 289 The CQT has a few characteristics that make it a better 290 fit for musical data when compared to the rapid Fourier 291 transform. Since the output of the transform is essentially 426 ultimate convolution layer, the classification layer outputs 427 a collection of confidence scores (a value between 0 and 428 1), that specifies how likely the image is to belong to a 429 class or set of the desired output. For example, if we have 430 a Network model that detects cats, dogs, and cars from just 431 their images, the output of the final layer may contain any 432 of those already considered input images. Figure 6 shows the 433 schematic diagram of the various layers in the CNN sequence, 434 which has been used to classify stacked spectrograms. 435 The pooling layer is used to reduce the spatial size of 436 the convolved features, like the convolutional layer. There 437 are mainly two styles of pooling: Average pooling and Max 438 pooling. In Max pooling, the maximum value of a pixel from 439 a part of the image covered by the kernel is selected. This 440 layer discards the noisy activation altogether and also per-441 forms de-noising alongside dimensionality reduction. On the 442 opposite hand, Average pooling returns the average of all 443 the values from the portion of the image covered by the 444 kernel. It generally performs dimensionality reduction, as a 445 noise suppressing mechanism, as a result of that, Max pooling 446 performs tons better than Average pooling. At the last stage 447 of the network, the fully connected layers are used, after 448 feature extraction and consolidation have been performed 449 by the convolutional and pooling layers. These are used to 450 create final non-linear combinations of features and then for 451 creating the final predictions by the network. 452 As we have already covered, CNNs do well when it comes 453 to picture classification. Additionally, earlier studies have 454 already demonstrated how well-known CNN architectures, 455 including AlexNet, VGG, Inception, and ResNet, performed 456 when applied to audio-based classifications. To create a spec-457 trogram, which served as an input to the network models, the 458 In this paper, we proposed transfer learning 2D-CNN models 822 to create an ensemble learning-based framework for cate-823 gorizing musical instruments by synthesizing a 3-channel 824 spectrogram. The extraction of significant features from spec-825 trograms of audio data has been demonstrated using models 826 pre-trained on large image datasets, effectively converting 827 audio processing and detection into a computer vision chal-828 lenge. The proposed MIC_FuzzyNET model combines trans-829 fer learning, CNNs, and a fuzzy rank-based ensemble tech-830 nique using the Gompertz function and a 3-channel stacked 831 spectrogram. Because the datasets used for training the deep 832 CNNs are not large, transfer learning is a good option. The 833 dynamic assignment of ranks to the classifiers allows pre-834 dictions to be produced for new datasets without the need to 835 initialize a new set of weights for the entire ensemble phase 836 875 instrument classification using fractional Fourier transform based-MFCC 876 features and counter propagation neural network,''

JOURNAL-AUDIO ENGINEERING SOCIETY, 2001

A study on the automatic classification of musical instrument sounds is presented. A database of musical instrument sounds parameters was built for this purpose, which consists of musical instrument recordings and their parametric representation. The parameterization process was conceived and performed in order to find significant musical instrument sound features and to remove redundancy from the musical signal. Classification experiments of musical instrument sounds were performed with neural networks allowing a discussion of the feature extraction process efficiency and of its limitations. Conclusions and remarks concerning further development of this study and its relation to the current MPEG-7 standardization process are included.

International Journal of Research in Advent Technology, 2019

Communication plays a vital role according to the people's emotion, as emotions and gesture play 80% role while communication. Nowadays emotion recognition and classification are used in different areas to understand the human feelings like in the robotics, Health care, Military, Home automation, Hands-free computing, Mobile Telephony, Video game,call-center system, Marketing, etc. SER can help better interaction between the machine and the human. There are various algorithms and combination of the algorithms are available to recognize and classify the audio according to their emotion. In this paper, we attempted to investigate the episodic significant works, their technique and the impact of the approaches and the scope of the correction of the results.

2004

Several factors affecting the automatic classification of musical audio signals are examined. Classification is performed on short audio frames and results are reported as "bag of frames" accuracies, where the audio is segmented into 23ms analysis frames and a majority vote is taken to decide the final classification. The effect of different parameterisations of the audio signal is examined. The effect of the inclusion of information on the temporal variation of these features is examined and finally, the performance of several different classifiers trained on the data is compared. A new classifier is introduced, based on the unsupervised construction of decision trees and either linear discriminant analysis or a pair of single Gaussian classifiers. The classification results show that the topology of the new classifier gives it a significant advantage over other classifiers, by allowing the classifier to model much more complex distributions within the data than Gaussian schemes do.